多值决策表的最小决策树生成  

Minimal Decision Tree Generation for Multi-Label Decision Tables

作　　者：乔莹 许美玲 钟发荣[1] 曾静[1] 莫毓昌[1] 

机构地区：[1]浙江师范大学,浙江金华

出　　处：《计算机科学与应用》2016年第10期617-628,共12页Computer Science and Application

基　　金：国家自然科学基金面上项目(61572442,61272130);浙江省科技厅公益性技术应用研究项目(2015C33085);浙江省教育厅项目(Y201226127)。

摘　　要：决策树技术在数据挖掘的分类领域应用极其广泛,可以从普通决策表(每行记录包含一个决策值)中挖掘有价值的信息,但是要从多值决策表(每行记录包含多个决策值)中挖掘潜在的信息则比较困难。多值决策表中每行记录包含多个决策值,多个决策属性用一个集合表示。针对已有的启发式算法,如贪心算法,由于性能不稳定的特点,该算法获得的决策树规模变化较大,本文基于动态规划的思想,提出了使决策树规模最小化的算法。该算法将多值决策表分解为多个子表,通过多值决策表的子表进行构造最小决策树,进而对多值决策表进行数据挖掘。Decision tree is a widely used classification in data mining. It can discover the essential knowledge from the common decision tables (each row has a decision). However, it is difficult to do data mining from the multi-label decision tables (each row has a set of decisions). In a multi-label decision tables, each row contains several decisions, and several decision attributes are represented using a set. By testing the existing heuristic algorithms, such as greedy algorithms, their performance is not stable, i.e., the size of the decision tree might become very large. In this paper, we propose a dynamic programming algorithm to minimize the size of the decision trees for a multi- label decision table. In our algorithm, the multi-label decision table is divided into several subtables, and the decision tree is constructed by using all subtables of the multi-label decision table, then useful information can be discovered from the multi-label decision tables.

关 键 词：多值决策表 决策树 动态规划算法 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]